Open in tt-awesome →

M-Stack

community
by MultiCoreWare · Python · Apache-2.0 · 22⭐ ·

MultiCoreWare's declarative Python SDK for building AI inference platforms on Kubernetes, from bare machines to a served model. Tenstorrent is a first-class accelerator type: the Helm layer installs the tt-operator chart and the serving layer exposes OpenAI-compatible endpoints over vLLM.

📦 Repo
# Changelog

Notable changes to this SDK. Format follows
[Keep a Changelog](https://keepachangelog.com/en/1.1.0/); versions follow
[Semantic Versioning](https://semver.org/spec/v2.0.0.html).

Nothing has been released yet, so everything below is unreleased and the
public surface is still free to move.

## [Unreleased]

### Added

- **The capability pattern.** The capability is the class and the
  implementation is data: `Storage(type="longhorn")` rather than
  `LonghornStorage`. `multistack/capability.py` holds the shared
  machinery — type validation, typed `options` matched against `type`,
  namespace defaulting, driver dispatch and caching — so a new capability
  declares its choices and its own fields rather than reimplementing any
  of it. `docs/structure.md` describes the shape and how to extend it.
- **`multistack.stack.Stack`**, a composition root. States which cluster
  once, explicitly, and threads it into every spec; publishes what each
  layer produced for the next (`record()`), and refuses to build a spec
  whose `REQUIRES` name a capability that has published nothing. Not
  ambient — nothing is read from the environment, and `stack.outputs`
  shows exactly what will be filled in. Specs remain usable standalone.
- **Declared dependencies.** Every spec carries `REQUIRES`:
  `RKE2Cluster ()`, `Storage ("cluster",)`,
  `MinIOTenant ("cluster", "storage")`, `VLLMService ("cluster",)`.
  Machine-readable, so a composer can order operations.
- **Volume claims** — `VolumeClaim` plus `StorageBackend.create_claim`,
  `resize_claim`, `delete_claim`. Implemented on the capability, not per
  driver, since a claim names a StorageClass and every implementation
  produces one. Refuses to shrink, refuses a class without
  `allowVolumeExpansion`, and waits for `Bound` by default.
- **A Helm layer** at `multistack/helm/`. `HelmRunner` is synchronous; the
  async dependency stops behind it, so no driver or example writes
  `await`. Ships `redact()`, atomic upgrades by default, chart-name
  resolution with a cached index, and a values check that reports unknown
  paths rather than refusing them.
- `HelmRunner.get_release()` — the same lookup as `status()`, returning
  `None` instead of raising when there is no such release. From the
  `helm-manager-sdk` branch (Nikhil Rai, 89869a7), reimplemented on the
  direct revision lookup rather than a full release listing.
- **Shared cluster primitives** in `multistack/kube.py`: `require_cluster`,
  `require_storage_class`, `require_cli`, `run_local`, `kubectl`, `apply`,
  `wait_for`, `wait_for_job`, `wait_for_phase`, `wait_for_absent`,
  `node_name_for`.
- `VLLMService.runtime_class`, needed for GPU serving on RKE2 before CDI —
  without `runtimeClassName: nvidia` the pod schedules and then finds no
  GPU.
- Guards for the things nothing was checking: every example and every
  ```python doc snippet must parse and its imports must resolve; markdown
  links and paths named in examples must exist; `SUPPORTED_TYPES`
kubernetes helm tt-operator vllm kserve inference sdk python multicoreware